A generative framework for predicting cellular morphological and transcriptomic perturbation responses.

Area of Science:

• Computational Biology
• Drug Discovery
• Cellular Imaging

Background:

• Predicting cellular responses to perturbations is vital for drug discovery and virtual cell modeling.
• Existing methods often predict single data types or need experimental data, limiting de novo screening.
• A unified approach is needed to predict multiple cellular responses simultaneously from perturbations.

Purpose of the Study:

• To introduce MultiVCDiff, a multimodal generative diffusion framework.
• To enable simultaneous prediction of high-fidelity cellular morphology images and transcriptomic profiles.
• To facilitate de novo virtual screening and accelerate drug discovery.

Main Methods:

• Developed a unified multimodal generative diffusion framework (MultiVCDiff).
• Trained the framework on a trimodal corpus of 1,118 chemical and 130 genetic perturbations.
• Evaluated performance in strict zero-shot settings on unseen drugs.

Main Results:

• MultiVCDiff outperforms state-of-the-art single-modality methods in zero-shot prediction.
• The framework successfully predicts both cellular morphology and transcriptomic profiles accurately.
• It effectively captures biological relationships and identifies drug mechanisms of action.

Conclusions:

• MultiVCDiff enables scalable in silico hypothesis generation by removing the need for post-perturbation data.
• The framework significantly accelerates drug discovery processes.
• It advances the development of holistic virtual cell models.